Pleiotropic loci underlying bone mineral density and bone size identified by a bivariate genome-wide association analysis

Abstract

Summary

Aiming to identify pleiotropic genomic loci for bone mineral density and bone size, we performed a bivariate GWAS in five discovery samples and replicated in two large-scale samples. We identified 2 novel loci at 2q37.1 and 6q26. Our findings provide insight into common genetic architecture underlying both traits.

Introduction

Bone mineral density (BMD) and bone size (BS) are two important factors that contribute to the development of osteoporosis and osteoporotic fracture. Both BMD and BS are highly heritable and they are genetically correlated. In this study, we aim to identify pleiotropic loci associated with BMD and BS.

Methods

We conducted a bivariate genome-wide association (GWA) analysis of hip BMD and hip BS in 6180 participants from 5 samples, followed by in silico replication in the UK Biobank study of BMD (N = 426,824) and the deCODE study of BS (N = 28,954), respectively.

Results

SNPs from 2 genomic loci were significant at the genome-wide significance (GWS) level (p lt; 5 × 10⁻⁸) in the discovery samples and were successfully replicated in the replication samples (2q37.1, lead SNP rs7575512, discovery p = 1.49 × 10⁻¹⁰, replication p = 0.05; 6q26, lead SNP rs1040724, discovery p = 1.95 × 10⁻⁸, replication p = 0.03). Functional annotations suggested functional relevance of the identified variants to bone development.

Conclusion

Our findings provide insight into the common genetic architecture underlying BMD and BS, and enhance our understanding of the potential mechanism of osteoporosis fracture.

Introduction

Bone, a major tissue type of human body, plays an important role in organism activity, such as protecting internal organs, maintaining body posture, and constituting motor system. Bone mineral density (BMD) is an important determinant for bone strength and fracture risk. The change of BMD can lead to severe bone-related diseases, such as osteoporosis and fracture. Though influenced by environmental factors, BMD is a highly heritable trait with heritability ranging from 50 to 80% [1].

Bone size (BS) is another important factor that influences bone geometry and bone strength. Many recent studies have suggested that hip bone size could be a useful measurement of hip fracture risk [2-4]. BS is predominantly determined by inheritance, with heritability as high as 75% [5].

Previous studies have reported some genes that show potential genetic interplay between BMD and BS, such as PLCL1 (phospholipase c-like 1) [6] and UQCC (ubiquinol-cytochrome reductase complex chaperone) [7]. Among them, PLCL1 was reported to be associated with bone density and hip bone size by two independent studies [6, 8], strengthening the confidence towards its pleiotropic effect to both BMD and BS.

To date, a large number of genome-wide association studies (GWASs) and their meta-analyses have been conducted for BMD, and hundreds of genomic loci have been identified [9-12]. For BS, a limited number of GWASs have been conducted, and specific genes underlying variation of BS are largely unknown. Furthermore, the current GWASs and their meta-analyses routinely analyze each trait individually. Joint consideration of correlated traits can provide additional information compared with information contained in individual traits, and could improve power to detect pleiotropic genes [13, 14].

Here, we conduct a bivariate GWAS meta-analysis of hip BMD and hip BS, aiming to identify pleiotropic loci jointly regulating both traits.

Materials and methods

Study samples

All samples were approved by the respective institutional ethics review boards, and all participants provided written informed consent. The details of the sample recruitment procedure, quality control, and statistical analysis method were described elsewhere [8]. Briefly, the first sample comprises 986 unrelated subjects from the Omaha Osteoporosis Study (OOS). The second sample comprises 2281 unrelated subjects from the Kansas City Osteoporosis Study (KCOS). The third sample comprises 1619 unrelated subjects of Chinese Han ancestry from the China Osteoporosis Study (COS). Both the fourth and fifth samples are from the Women’s Health Initiative (WHI) observational study that was accessed through the dbGAP. The WHI is a partial factorial randomized and longitudinal cohort with > 160,000 women aged 50–79 years of diverse ancestries [15]. The GWAS was conducted in two minority populations of African-American ancestry and Hispanic ancestry, including ~ 12,000 participants. Of them, ~ 10% participants received dual-energy X-ray absorptiometry (DEXA) scan. After checking the availability of both genotypes and phenotypes, we identified the fourth sample comprising 843 subjects of African-American ancestry (WHI-AA) and the fifth sample comprising 446 subjects of Hispanic ancestry (WHI-HIS).

Phenotype measurements

BMD and BS were measured at proximal femur of hip by scanners following the manufacturer protocols (Hologic Inc., Bedford, MA, USA). Covariates (including gender, age, age squared, and height) were screened for significance with the step-wise linear regression model. To adjust for potential population stratification, the first five principal components derived from genome-wide genotype data were included as covariates. Raw BMD and BS values were adjusted by significant covariates, and the residuals were normalized by inverse quantiles of standard normal distribution.

Genotyping and quality control

All GWAS samples were genotyped by high-throughput SNP genotyping arrays (Affymetrix Inc., Santa Clara, CA, USA; or Illumina Inc., San Diego, CA, USA within individual samples), following the manufacturer’s protocols. Quality control (QC) within each sample was implemented at both the individual and SNP level. At the individual level, sex compatibility was checked by imputing sex from X-chromosome genotype data with PLINK [16]. Individuals of ambiguous imputed sex or of imputed sex inconsistent with reported sex were removed. At the SNP level, SNPs violating the Hardy-Weinberg Equilibrium (HWE) rule (p value < 1.0 × 10⁻⁵) were removed. Population outliers were monitored by genotype-derived principle components, and outliers were removed.

Genotype imputation

All GWAS samples were imputed by the 1000 genomes project phase 3 sequence variants (as of May 2013) [17]. Haplotypes representing 240 individuals of European ancestry, 244 of East Asian ancestry, 319 of African ancestry, and 170 of admixed American ancestry were downloaded from the project download site. Haplotypes of bi-allelic variants, including SNPs and bi-allelic insertions/deletions (indels), were extracted to form reference panels for imputation. As a QC procedure, variants with zero or one copy of minor allele were removed.

Each GWAS sample was imputed by the respective reference panel with the closest ancestry. Prior to imputation, a consistency test of allele frequency between the GWAS and reference samples was examined with the chi-square test. To correct for potential mis-strandedness, GWAS SNPs that failed the consistency test (p lt; 1.0 × 10⁻⁶) were transformed into the inverse strand. SNPs that again failed the consistency test were removed from the GWAS sample. Imputation was performed with FISH [18], a fast and accurate diploid genotype imputation algorithm that we previously developed.

Association analysis in individual studies

Each GWAS sample was tested for association between normalized phenotype residuals and genotyped and imputed genotypes under an additive model of inheritance. Both univariate and bivariate association tests were performed in each individual sample. The association was examined by the univariate/bivariate linear regression model. Univariate associations were examined with MACH2QTL [19], and bivariate associations were examined by our in-house JAVA software [20].

Meta-analysis

Summary association statistics from individual GWAS samples were combined to perform univariate or bivariate meta-analysis. As a QC step, only well-imputed and common or less common (minor allele frequency, MAF ;gt;0.01 in the European population) SNPs were included in the analysis. Well-imputed SNPs were defined as SNPs whose imputation certainty measure r² > 0.3 in at least 2 samples.

Both univariate and bivariate meta-analysis were performed under the fixed-effects model [21]. Briefly, for a particular SNP, let β_i, = (β_1i, β_2i) be the vector of the regression coefficient for the two traits in the ith study (i = 1, …, n, n = 5) and let ${\mathbf{V}}_i=\left[\begin{array}{cc}\hfill v_{11}\hfill & \hfill v_{12}\hfill \\ \hfill v_{21}\hfill & \hfill v_{22}\hfill \end{array}\right]$ be the corresponding symmetric variance-covariance matrix for the two regression coefficients. Both β_i and V_i are obtained from individual study analysis. Define the following data structure

$$\begin{gathered} \mathbf{B}={\left({\beta }_1^{\prime },{\beta }_2^{\prime },\dots ,{\beta }_n^{\prime }\right)}_{2n\times 1}^{\prime } \\ x=\left[\begin{array}{ccc}\hfill 1\hfill & \hfill \hfill & \hfill 0\hfill \\ \hfill 0\hfill & \hfill \hfill & \hfill 1\hfill \\ \hfill \hfill & \hfill \cdots \hfill & \hfill \hfill \\ \hfill 1\hfill & \hfill \hfill & \hfill 0\hfill \\ \hfill 0\hfill & \hfill \hfill & \hfill 1\hfill \end{array}\right],v={\left[\begin{array}{cccc}{v}_1\hfill & \hfill & \hfill & \hfill \\ \hfill & v_2\hfill & \hfill & 0\hfill \\ \hfill & \hfill & \cdots \hfill & \hfill \\ \hfill & 0\hfill & \hfill & v_n\hfill \end{array}\right]}_{2n\times 2} \end{gathered}$$

where B is the vector of the regression coefficient, X is the design matrix, and V is the variance-covariance matrix for all studies, respectively.

The generalized least-squared estimator $\widehat{\beta }$ of overall regression coefficients is given by

$$\widehat{\beta }={\left({\mathbf{X}}^{{}^{\prime }}{\mathbf{V}}^{-1}\mathbf{X}\right)}^{-1}{\mathbf{X}}^{{}^{\prime }}{\mathbf{V}}^{-1}\mathbf{B}$$

which has a normal distribution with mean β and covariance matrix Σ given by

$$\Sigma ={\left({\mathbf{X}}^{{}^{\prime }}{\mathbf{V}}^{-1}\mathbf{X}\right)}^{-1}$$

Define the bivariate score statistic as BT, which has the form

$$BT={\widehat{\beta }}^{{}^{\prime }}{\Sigma }^{-1}\widehat{\beta }$$

Under the null hypothesis of no association to either phenotype, that is, β = 0 (for both traits), BT will asymptotically follow a chi-squared distribution with 2 degrees of freedom.

Similarly, define the two univariate score statistics as UT₁ and UT₂, respectively, which have the forms

$$U{T}_1=\frac{{\widehat{\beta }}_{{}_1}^{{}^2}}{{\Sigma }_{{}_{11}}},U{T}_2=\frac{{\widehat{\beta }}_{{}_2}^{{}^2}}{{\Sigma }_{{}_{22}}}$$

where ${\widehat{\beta }}_1$ and ${\widehat{\beta }}_2$ are two elements in $\widehat{\beta }$, and Σ₁₁ and Σ₂₂ are two diagonal elements in Σ. Under the null hypothesis of no univariate association, that is, β₁ = 0 or β₂ = 0, UT₁ or UT₂ will follow a chi-squared distribution with 1 degree of freedom.

The above meta-analysis model was implemented in an in-house java program BiMeta.jar, which is available upon request to the corresponding authors.

Replication samples

John A. Morris et al. [11] recently reported the largest GWAS of BMD as estimated by quantitative ultrasound of the heel (eBMD) in 426,824 individuals from the UK Biobank. The summary results are publicly available at the GEFOS website (http://www.gefos.org/?q=documents). We downloaded the summary statistics and performed in silico replication analysis of BMD trait in the UK Biobank study.

For BS, Unnur Styrkarsdottir et al. [22] conducted a GWAS in 28,954 Icelandic participants. BS was measured at 5 skeletal sites including hip using DEXA scanners (Hologic). The GWAS summary statistics are available at https://www.decode.com/summarydata/. We downloaded the summary statistics and performed in silico replication for BS.

With GWAS summary statistics of the two replication samples, we also derived a bivariate p value using the CPASSOC algorithm [23].

Functional annotation

First, we performed cis-expression quantitative trait locus (cis-eQTL) analyses in bone-related tissues/cells, including human osteoclast cells, lymphoblastoid cells, and macrophage cells [24-26]. Significance level was set to be 0.05.

Second, we annotated the functional relevance of the identified SNPs with HaploReg v4.1 [27]. HaploReg annotates SNPs into different functional categories according to the information from a variety of large experiment projects. These categories include conservation sites, DNAse hypersensitivity region, transcription factor binding sites, promoter, enhancer, and others. We annotated lead SNPs and their neighbor SNPs with strong LD pattern (r² > 0.8).

Third, we defined candidate gene as the gene that is closest to the identified lead SNP. We constructed the protein–protein interaction (PPI) network of the identified candidate genes with STRING [28]. STRING uses information based on gene co-expression, text-mining, and others.

Lastly, we explored the functional relevance of the candidate genes in MGI mouse model database (Mouse Genome Informatics, http://www.informatics.jax.org/). This database reports gene mutations that are associated with bone-related outcomes.

Results

The basic characteristics of the discovery samples are listed in Table 1. A total of 6180 subjects from five samples are included in the meta-analysis.

Table 1.

The basic characteristics of the discovery samples

Sample	Source	Anc	N	Female (%)	Age	Weight (kg)	Height (cm)	Hip-BMD	Hip-BS	Bone densitometer
OOS	In-house	EUR	986	50.20	50.17(± 18.34)	80.11(± 17.73)	171.00(± 10.00)	0.97(± 0.16)	39.45(± 6.58)	Hologic
KCOS	In-house	EUR	2281	75.69	51.33(± 13.74)	75.26(± 17.53)	166.35(± 8.47)	0.97(± 0.18)	38.35(± 6.37)	Hologic
COS	In-house	EAS	1619	50.71	34.47(± 13.24)	60.09(± 10.48)	164.25(± 8.17)	0.92(± 0.13)	34.06(± 5.71)	Hologic
WHI-AA	dbGAP	AMR	843	100.00	61.17(± 7.30)	78.78(± 16.91)	162.75(± 5.77)	0.94(± 0.15)	35.00(± 3.55)	Hologic
WHI-HIS	dbGAP	AFR	446	100.00	60.09(± 7.52)	72.81(± 14.77)	158.16(± 5.55)	0.85(± 0.13)	32.92(± 2.78)	Hologic

The numbers within parentheses are standard deviation (SD)

Anc ancestries of the sample population, N the number of the samples, OOS Omaha osteoporosis study, KCOS Kansas City Osteoporosis Study, COS China Osteoporosis Study, WHI-HIS Hispanic ancestry of the Women’s Health Initiative observational study, WHI-AA African-American ancestry of the Women’s Health Initiative observational study

Discovery sample meta-analysis

The logarithmic quantile–quantile plot of meta-analyses shows a marked deviation in the tail of the distribution, implying the possible existence of true associations (Fig. 1).

Fig. 1.

Logarithmic quantile–quantile (QQ) plot of the discovery GWAS results. Results were plotted for univariate GWAS of hip BMD (dodgerblue), univariate GWAS of hip BS (olivedrab), and bivariate GWAS of both traits (darkgoldenrod). p1 is the p value of univariate hip BMD GWAS meta-analysis, p2 is the p value of univariate hip BS GWAS meta-analysis, and p is the p value of bivariate GWAS meta-analysis

Univariate GWAS meta-analysis identified 4 variants associated with BMD at the genome-wide significance (GWS, 5 × 10⁻⁸) level, mapping to 2 distinct loci: 12q13.11 (lead SNP rs139681378, p = 1.70 × 10⁻⁸) and 17q21.32 (lead SNP rs890431, p = 1.90 × 10⁻⁸). Here, an independent locus is defined as a genomic region of 500 kb on either side of the lead SNP. Both loci were reported to be associated with BMD in previous studies [10, 29], with different lead SNPs. The LD between the present lead SNP and the previously reported lead SNP is low (r² < 0.1) for both loci, indicating that the newly identified associations are independent of previous signals.

Univariate analysis of BS identified 31 variants at the GWS level, mapping to 27 distinct loci (Supplementary Table 1). However, most of them are heterogeneous (I² > 50%). Sensitivity analysis shows that the heterogeneity in all cases comes from the KCOS samples, which is extremely significant while none of the other samples is significant.

Bivariate GWAS meta-analysis identified 31 SNPs located at 26 distinct loci at the GWS level. Among the 29 loci identified by univariate analyses (2 for BMD and 27 for BS), 22 loci remain significant at the GWS level in the bivariate analysis. The bivariate analysis identified additional 4 loci that are not significant by either univariate analysis at the GWS level: 2p22.1 (rs143571077, p = 4.45 × 10⁻⁸), 6q26 (rs1040724, p = 1.95 × 10⁻⁸), 9q32 (rs2274784, p = 2.39 × 10⁻⁸), and 9q34.11 (rs2296793, p = 9.66 × 10⁻⁹). All these 4 loci are nominally significant in both univariate analyses, implying that these loci exert pleiotropic effect.

Among the 26 bivariately significant loci, 23 have been reported to be associated with BMD by several previous studies, while the remaining 3 loci are novel to both BMD and BS: 3q22.1 (rs1389271, p = 9.64 × 10⁻¹²), 7p13 (rs2289057, p = 5.55 × 10⁻¹²), and 12q12 (rs2653772, p = 1.31 × 10⁻⁸).

In total, bivariate analysis identified 26 loci that are nominally significant for both phenotypes. These loci imply a pleiotropic effect to both phenotypes. Manhattan plots of the univariate and bivariate GWAS meta-analyses are displayed in Fig. 2. All the bivariate association results are listed in Supplementary Table 2.

Fig. 2.

Manhattan plots of the meta-analyses for hip BMD, hip BS, and both traits from top to bottom, respectively. p1 is the p value of univariate hip BMD GWAS meta-analysis, p2 is the p value of univariate hip BS GWAS meta-analysis, and p is the p value of bivariate GWAS meta-analysis. The X-axis represents chromosome 1–22, and the Y-axis represents the log converted p value

Replication in the summary results

We replicated our bivariately significant lead SNPs in the UK Biobank GWAS summary results of eBMD [11] and in the deCODE GWAS summary results of BS [22]. Successful replication was defined as consistent effect direction as well as nominally significant p value (p < 0.05).

Of the 26 lead SNPs, rs7575512 at 2q37.1 (discovery p = 1.49 × 10⁻¹⁰) and rs1040724 at 6q26 (discovery p = 1.95 × 10⁻⁸) are successfully replicated in that: 1, both bivariate p values are nominally significant (p = 0.05 and 0.03); and 2, the effect direction is consistent between the discovery and replication samples for both individual traits (Table 2). Among the remaining SNPs, 5 are bivariately significant (p < 0.05); however, the effect direction is not consistent for either trait (Supplementary Table 3).

Table 2.

Main results of the identified SNPs

SNP	Chr	Position	Locus	EA/ OA	Closest gene	Discovery sample					Replication sample
						BMD		BS		Bivariate	UKB for BMD		deCODE for BS		Bivariate
						Beta	p	Beta	p	p	Beta	p	Beta	p	p
rs7575512	2	235210293	2q37.1	T/C	ARL4C	0.049	0.02	0.132	3.26 × 10−10	1.49 × 10−9	0.009	0.06	0.045	0.03	0.05
rs1040724	6	161575543	6q26	T/C	AGPAT4	0.115	4.01 × 10−5	0.140	3.95 × 10−7	1.95 × 10−8	0.008	0.05	0.005	0.75	0.03

Physical position is based on the human genome GRCH37 assembly

Chr chromosome, EA effect allele, OA other allele, Beta the regression coefficient of the effect allele, Closest gene the closest gene to which the SNP mapped

Function annotation

We annotated the two identified lead SNPs rs7575512 and rs1040724 from various sources of information.

We first looked up their eQTL activities in several bone-related cells. Both SNPs have eQTL activities in more than one cell line. Specifically, rs7575512 is associated with ACVR1B (Activin A Receptor Type 1B) gene expression in macrophages (p = 7.50 × 10⁻⁶) [24] and ITGB1 (Integrin Subunit Beta 1) gene expression in lymphoblastoid (p = 8.00 × 10⁻⁵) [25]. rs1040724 is associated with PTK2 (Protein Tyrosine Kinase 2) gene expression in lymphoblastoid (p = 1.00 × 10⁻⁴) [25] and AGPAT4 (1-acylglycerol-3-phosphate O-acyltransferase 4) gene expression in thyroid tissue (p = 6.30 × 10⁻¹⁹) [30]. All the three genes (ACVR1B, ITGB1 and PTK2) may play a role in bone development. For example, gene ACVR1B is associated with abnormal bone structure in a mouse model experiment [31]. ITGB1 and PTK2 are in the same one protein–protein interaction network based on the result from STRING website. They are also connected with another gene TLN1 (Talin 1), which is reported to be associated with eBMD and body height [10].

We then explored their functional relevance with the online tool HaploReg v4.1 [27]. rs7575512 has enhancer activity in 10 tissues and DNase activity in blood. Among these tissues are osteoblast primary cells, peripheral blood monocytes (PBMs), and B and T lymphocytes that are related to bone metabolism [32]. rs1040724 itself does not have regulatory activity. But two of its neighboring SNPs with strong LD (r² > 0.8), rs77901568 and rs7751159, have enhancer activity in osteoblast primary cells, PBMs, and B and T lymphocytes.

rs7575512 is located 191 kb from the 3′- of ARL4C gene (ADP-ribosylation factor-like 4C), which is closest to the SNP and is defined as candidate gene. rs1040724 is located in an intronic region of gene AGPAT4. We explored the PPI network of both candidate genes with STRING. ARL4C is tightly connected with RPL31 (Ribosomal Protein L31) (Fig. 3), which is reported to be associated with osteoporosis and lipid metabolism [33, 34]. AGPAT4 is directly associated with GPAT2 (glycerol-3-phosphate acyltransferase 2, mitochondrial) (Fig. 4), which is reported to play a role in the variation of eBMD previously [35].

Fig. 3.

The gene interaction network of ARL4C was downloaded from the STRING database. Colors of edges refer to the type of evidence linking the corresponding proteins: red, gene fusion; dark blue, co-occurrence; black, coexpression; magenta, experiments; cyan, databases; light green, text mining; mauve, homology. From the figure, we could get a message that ARL4C was co-expression with RPL3

Fig. 4.

The gene interaction network of AGAPT4 was downloaded from the STRING database. Colors of edges refer to the type of evidence linking the corresponding proteins: red, gene fusion; dark blue, co-occurrence; black, coexpression; magenta, experiments; cyan, databases; light green, text mining; mauve, homology. From the figure, we could get a message that AGAPT4 was tightly associated with GPAT2

At last, we explored the function of both candidate genes and their related genes with bone metabolism in the MGI mouse model database. ARL4C is not found to be associated with bone traits. It is a member of the ADP-ribosylation factor family of GTP-binding proteins. Another member of this family, ARL4D, is associated with the decreased BMD and BMC (bone mineral content) [36, 37], implying the relevance of ARL4C. AGPAT4 is associated with decreased BMC and lean body mass according to a previous mouse model experiment study [31].

Discussion

In this study, we have conducted a bivariate genome-wide association meta-analysis and replicated the significant results in the UK Biobank and deCODE GWAS summary data. We identified 31 SNPs located in 26 loci bivariately associated with BMD and BS and replicated two loci 2q37.1 and 6q26.

rs7575512 is located in the region 2q37.1, in which several SNPs are reported to be associated with BMD, i.e., rs838721, rs2675952, and rs73102769 [9, 10, 29]. These SNPs are located in intronic region of gene DGKD (diacylglycerol kinase delta) and NGEF (neuronal guanine nucleotide exchange factor), respectively. However, the LD between these SNP and rs7575512 is weak (r² < 0.1), implying an independent signal of the previously identified ones. rs7575512 has enhancer activity in osteoblast primary cells, peripheral blood monocytes (PBMs), and B and T lymphocytes. PBMs have been well established as a working cell model for studying gene expression patterns in relation to osteoporosis risk [38]. They may act as precursors of osteoclasts since they can differentiate into osteoclasts [39], and they express different cytokines which are important for osteoclast differentiation, activation, and apoptosis [40]. B lymphocytes, an important immune cell type, express/secrete factors involved in osteoclastogenesis, such as receptor tumor necrosis factor superfamily member 11 and osteoprotegerin [41].

The PPI network analysis suggested that ARL4C is connected with RPL31, where the latter is reported to participate in the process of lipid metabolism [33]. LDL-c, one type of lipid, is involved in the formation and extension of tartrate-resistant acid phosphatase multinucleated cells (osteoclast-like cells), which are abrogated with the depletion of LDL. Another supporting evidence is that oxidized LDL may boost osteoclast differentiation through inducing osteoclast-associated receptor in endothelial cells.

The other identified lead SNP rs1040724 is located in an intronic region of gene AGPAT4 at 6q26. Its eQTL activity in thyroid tissue implies its association with hormone-related disorders. Thyroxine profoundly alters bone turnover by a direct action on the bone cells and by influencing calcium compartment sizes as well as flowing to and from these compartments [42]. Furthermore, subclinical hyperthyroidism causes a reduction in BMD and increases fracture risk, in particular in association with estrogen deficiency [43]. Given the tight link between thyroid hormone and bone metabolism, the relationship between AGAPT4 and the homeostasis of thyroid hormones may be worthy of a further investigation.

Certain limitations exist in this study. First, the discovery samples consist of five samples of diverse ancestries. It is well known that different ancestries may have different LD structures and genetic effects. Second, as BMD and BS can derive each other in a single formula, the two measures may be co-linear. We kept in mind that their relationship is not linear. For example, the Pearson correlation coefficient is only 0.33 in the largest KCOS sample. This moderate level of phenotypic correlation indeed motivated the analysis of two phenotypes jointly as the work done in the present study. Third, while the discovery samples used hip BMD measured by the DEXA scan, the replication UKB sample used estimated heel BMD. Previous studies have shown that the genomic loci identified by eBMD GWAS were largely overlapped (84%) with those identified by DEXA-derived BMD GWAS [9, 11]. These evidences indicated that eBMD could partially replicate the findings identified in other skeletal sites including hip, though the two traits are not perfectly matched.

In conclusion, by conducting a bivariate genome-wide association meta-analysis study followed by in silico replication, we identified 2 novel pleiotropic loci 2q37.1 and 6q26 for BMD and BS. Further research effort is warranted to uncover the functional mechanism behind the genetic association. Our findings provide insights into the common genetic architecture underlying BMD and BS and enhance our understanding of the potential mechanism of osteoporotic fracture.